search

LSSTDESC / descwl_coadd

Tools for coadding using the LSST DM software stack
GNU General Public License v3.0
2 stars 2 forks source link

deal with nan in coadd #12

Closed esheldon closed 4 years ago

esheldon commented 4 years ago

there can be cases where there are no exposures over part of the coadd

make sure that no nan or other sentinels are being sent to metacal

esheldon commented 4 years ago

may want to deal with this in metadetect

beckermr commented 4 years ago

This shouldn't happen for our sims. Can you plot some images for one or point me to a case where this happened?

esheldon commented 4 years ago

what I saw was nan in the variance of the coadd, but I think this implies nan in the coadd or some other sentinel

esheldon commented 4 years ago

with the rotations it is plausible there is no information in some part of the final coadd

beckermr commented 4 years ago

Ahhhhh I forgot. So I make the SE observations sqrt(2) bigger than the coadd. This doesn't account for shear, but for realistic shear levels in the WCS, the central 300 pixels should always be covered. I think we should cut down to those. This is how I do my own sims.

esheldon commented 4 years ago

I think we should leave in the possibility of such gaps because in real data they will exist.

I think the coadd code does something sensible with this, so probably it is downstream when we send the images to metacal that we need to be careful.

esheldon commented 4 years ago

I checked, the image also gets nan

beckermr commented 4 years ago

Do they set mask bits?

esheldon commented 4 years ago

probably, I've not dealt at all with mask stuff yet

esheldon commented 4 years ago

in #3 I am currently setting the image and noise image to zero where the variance is zero (turning nan to 0)

Might want to replace with noise